Virus–Virus Interactions Impact the Population Dynamics of Influenza and the Common Cold

Virus–virus interactions impact the population dynamics of influenza and the common cold

Sema Nickbakhsha, Colette Maira,b, Louise Matthewsc,1, Richard Reevec, Paul C. D. Johnsonc, Fiona Thorburnd, Beatrix von Wissmanne, Arlene Reynoldsf, James McMenaminf, Rory N. Gunsong, and Pablo R. Murciaa,1

aMRC-University of Glasgow Centre for Virus Research, Institute of Infection, Immunity and Inflammation, College of Medical, Veterinary and Life Sciences, University of Glasgow, G61 1QH Glasgow, United Kingdom; bSchool of Mathematics and Statistics, College of Science and Engineering, University of Glasgow, G12 8QQ Glasgow, United Kingdom; cBoyd Orr Centre for Population and Ecosystem Health, Institute of Biodiversity, Animal Health and Comparative Medicine, College of Medical, Veterinary and Life Sciences, University of Glasgow, G12 8QQ Glasgow, United Kingdom; dThe Queen Elizabeth University Hospital, NHS Greater Glasgow and Clyde, G51 4TF Glasgow, United Kingdom; ePublic Health, NHS Greater Glasgow and Clyde, G12 0XH Glasgow, United Kingdom; fHealth Protection Scotland, NHS National Services Scotland, G2 6QE Glasgow, United Kingdom; and gWest of Scotland Specialist Virology Centre, NHS Greater Glasgow and Clyde, G31 2ER Glasgow, United Kingdom

Edited by Burton H. Singer, University of Florida, Gainesville, FL, and approved November 12, 2019 (received for review June 27, 2019) The human respiratory tract hosts a diverse community of cocirculating influenza A virus (IAV) pandemic of 2009 further galvanized viruses that are responsible for acute respiratory infections. This interest in the epidemiological interactions among respiratory shared niche provides the opportunity for virus–virus interac- viruses. It was postulated that rhinovirus (RV) may have delayed tions which have the potential to affect individual infection risks the introduction of the pandemic virus into Europe (12, 13), and in turn influence dynamics of infection at population scales. while the pandemic virus may have, in turn, interfered with ep- However, quantitative evidence for interactions has lacked suit- idemics of respiratory syncytial virus (RSV) (14, 15). Clinical able data and appropriate analytical tools. Here, we expose and studies utilizing coinfection data from diagnostic tests have also quantify interactions among respiratory viruses using bespoke suggested virus–virus interference at the host scale (13, 16, 17). analyses of infection time series at the population scale and coin- The role of adaptive immunity in driving virus interferences fections at the individual host scale. We analyzed diagnostic data that alter the population dynamics of antigenically similar virus from 44,230 cases of respiratory illness that were tested for 11 tax- strains is well known (18, 19). For example, antibody-driven onomically broad groups of respiratory viruses over 9 y. Key to our cross-immunity is believed to restrict influenza virus strain di- analyses was accounting for alternative drivers of correlated infec- versity, leading to sequential strain replacement over time (20). POPULATION BIOLOGY tion frequency, such as age and seasonal dependencies in infection Such antibody-driven virus interactions might even shape the risk, allowing us to obtain strong support for the existence of neg- temporal patterns of RSV, human parainfluenza virus (PIV), ative interactions between influenza and noninfluenza viruses and and human metapneumovirus (MPV) infections, which are tax- positive interactions among noninfluenza viruses. In mathematical onomically grouped into the same virus family (21). However, simulations that mimic 2-pathogen dynamics, we show that tran- where antigenically distant virus groups are concerned, the mech- sient immune-mediated interference can cause a relatively ubiqui- anisms are obscure, although the variety of possibilities in- tous common cold-like virus to diminish during peak activity of a cludes innate immunity, resource competition, and other cellular seasonal virus, supporting the potential role of innate immunity in driving the asynchronous circulation of influenza A and rhinovirus. Significance These findings have important implications for understanding the linked epidemiological dynamics of viral respiratory infections, an When multiple pathogens cocirculate this can lead to compet- important step towards improved accuracy of disease forecasting itive or cooperative forms of pathogen–pathogen interactions. models and evaluation of disease control interventions. It is believed that such interactions occur among cold and flu viruses, perhaps through broad-acting immunity, resulting in epidemiology | virology | ecology interlinked epidemiological patterns of infection. However, to date, quantitative evidence has been limited. We analyzed a he human respiratory tract hosts a community of viruses that large collection of diagnostic reports collected over multiple Tcocirculate in time and space, and as such it forms an eco- years for 11 respiratory viruses. Our analyses provide strong logical niche. Shared niches are expected to facilitate interspecific statistical support for the existence of interactions among re- interactions which may lead to linked population dynamics among spiratory viruses. Using computer simulations, we found that distinct pathogen species (1, 2). In the context of respiratory in- very short-lived interferences may explain why common cold fections, a well-known example is the coseasonality of influenza infections are less frequent during flu seasons. Improved un- and pneumococcus, driven by an enhanced susceptibility to sec- derstanding of how the epidemiology of viral infections is ondary bacterial colonization subsequent to influenza infection (3, interlinked can help improve disease forecasting and evalua- 4). Indeed, respiratory bacteria–bacteria and bacteria–virus inter- tion of disease control interventions. actions, and the underlying mechanisms by which they arise, are extensively studied (5–8). In contrast, evidence for the occurrence Author contributions: S.N., L.M., R.R., P.C.D.J., R.N.G., and P.R.M. designed research; S.N., of virus–virus interactions remains scarce and the potential C.M., and P.C.D.J. performed research; F.T., A.R., J.M., and R.N.G. contributed new re- agents/analytic tools; S.N., C.M., and P.C.D.J. analyzed data; and S.N., C.M., L.M., R.R., mechanisms are elusive, demanding greater research attention. P.C.D.J., F.T., B.v.W., A.R., J.M., R.N.G., and P.R.M. wrote the paper. The occurrence of such interactions may have profound economic The authors declare no competing interest. implications, if the circulation of one pathogen enhances or di- This article is a PNAS Direct Submission. minishes the infection incidence of another, through impacts on This open access article is distributed under Creative Commons Attribution License 4.0 the healthcare burden, public health planning, and the clinical (CC BY). management of respiratory illness. 1To whom correspondence may be addressed. Email: [email protected] or Early interest in the potential for interference among taxo- [email protected]. nomically distinct groups of respiratory viruses stemmed from This article contains supporting information online at https://www.pnas.org/lookup/suppl/ epidemiological observations of the temporal patterns of infec- doi:10.1073/pnas.1911083116/-/DCSupplemental. tion in respiratory virus outbreaks (9–11). More recently, the

www.pnas.org/cgi/doi/10.1073/pnas.1911083116 PNAS Latest Articles | 1of9 Downloaded by guest on September 26, 2021 processes (22–25). Recent experimental models of respiratory vi- A100 rus coinfections have demonstrated several interaction-induced effects, from enhanced (26) or reduced (22, 23) viral growth to 75 the attenuation of disease (23, 24). It has also been shown that cell fusion induced by certain viruses may enhance the replication of 50 others in coinfections (26). However, despite epidemiological, clinical, and experimental indications of interactions among respiratory viruses, quantita- 25

tively robust evidence is lacking. Such evidence has been hard to Infection status (%) 0 acquire, due to a paucity of consistently derived epidemiological Jan Jul JanJulJulJan JanJul Jan JulJan JulJan Jul Jan Jul Jan Jul time-series data over a sufficient time frame to provide well- 05 06 07 08 09 10 11 12 13 powered analyses, and the lack of analytical approaches to dif- 100 B RV ferentiate true virus–virus interactions from other drivers of IAV IBV coseasonality, such as age-dependent host mixing patterns (27, 75 RSV 28) and environmental factors associated with virus survival and CoV – – AdV transmission (29 31). As a result, studies evaluating virus virus 50 MPV interactions from epidemiological time series have thus far not PIV3 PIV1 accounted for such alternative drivers of coseasonality. PIV4 Here, we apply a series of statistical approaches and provide 25 PIV2

robust statistical evidence for the existence of interactions among prevalence Relative 0 respiratory viruses. We examined virological diagnostic data 05 06 07 08 09 10 11 12 13 from 44,230 episodes of respiratory illness accrued over a 9-y Year time frame in a study made possible by the implementation of multiplex-PCR methods in routine diagnostics that allow the Fig. 1. Temporal patterns of viral respiratory infections detected among simultaneous detection of multiple viruses from a single respi- patients in Glasgow, United Kingdom, 2005 to 2013. (A) Percentage of pa- ratory specimen. Each patient was tested for 11 virus groups (28, tients diagnosed with a single viral infection (white), a viral coinfection 29), providing a single, coherent data source for the epidemio- (gray), or determined to be virus-negative (black) by multiplex RT-PCR in each calendar month from 2005 to 2013 (6-mo intervals depicted by vertical logical examination of infection dynamics of both cocirculating lines; Jan = January, Jul = July). (B) Relative virus prevalences in each cal- viruses in general and coinfection patterns in individual patients. endar month, from 2005 to 2013; note total virus counts may sum to more Using these data, our study addresses the following questions: 1) than those informing single infection prevalences due to coinfections, and Is there statistical evidence of virus–virus interactions in the test frequency denominators vary slightly across viruses. During the first temporal patterns of infection at the population scale?, 2) Is wave of the United Kingdom’s influenza A pandemic [A(H1N1)pdm09] in there statistical evidence of virus–virus interactions in coin- 2009, infections with influenza A virus were relatively more prevalent among fection patterns at the individual host scale?, and 3) Is a short- the patient population than noninfluenza virus infections (highlighted by black box). RV = rhinoviruses (A–C); IAV = influenza A virus (H1N1 and H3N2); lived immune-mediated interference at the scale of individual = = = hosts sufficient to induce asynchronous seasonal patterns of in- IBV influenza B virus; RSV respiratory syncytial virus; CoV human coronaviruses (229E, NL63, HKU1); AdV = human adenoviruses; MPV = human fection at the population scale? metapneumovirus; PIV3 = parainfluenza 3 virus; PIV1 = parainfluenza 1 virus; PIV4 = parainfluenza 4 virus; PIV2 = parainfluenza 2 virus. See also Table 1. Results Virus groups are listed in descending order of their total prevalence. The Overall Prevalence of Any Viral Respiratory Infection among Patients with Respiratory Illness Is Relatively Stable over Time, Despite Strongly Varying Prevalences of Individual Viruses. We first evaluated ever, examining whether these are genuine virus–virus interac- the total monthly infection prevalences across all viral respiratory tions (mediated at either the host or population level) required infections from 2005 to 2013. As typically observed in tem- us to address several methodological limitations in this relatively perate regions, the proportion of patients with respiratory ill- simple approach: It fails to account for autocorrelation in the ness testing positive to at least one respiratory virus peaked time series of individual viruses, or for potentially confounding during winter, with the exception of the influenza A H1N1 factors which might independently explain correlations, and it pandemic in the summer of 2009 (Fig. 1A). Nevertheless, even can produce spurious negative correlations with proportional during the influenza pandemic, the overall viral infection data or, alternatively, spurious positive correlations with absolute prevalence among patients remained broadly stable due to a infection counts. simultaneous decline in the contribution of noninfluenza viruses Traditional analytical methods are unable to address all of B to the total infection burden (Fig. 1 ). Throughout the 9-y these limitations simultaneously, so we developed an approach study period, because of seasonal fluctuations in the magni- that extends a multivariate Bayesian disease-mapping framework tude and timing of peaks in prevalences of individual viruses to infer interactions between virus pairs (32). This framework (Fig. 2), the dominating virus varied on a month-by-month basis estimates pairwise correlations by modeling observed monthly (Fig. 1B). virus counts relative to what would be expected in each month. Respiratory Viruses Exhibit Cross-Correlations at the Population Level Patient covariates age, gender, and general practice versus hos- That Are Independent of Seasonality. We evaluated correlations in pital origin (as a proxy for illness severity) were used to estimate the monthly prevalence time series for each pair of respiratory expected counts within each month for each virus independently, viruses. We first employed a simple bivariate nonparametric capturing age and typical seasonal variability in infection risk. For cross-correlation analysis by estimating Spearman’s rank corre- example, viral exposure events may be seasonally (anti-) correlation coefficients and observed 26 significant virus–virus correlated due to similarities (differences) in the climatic preferences of lations, 13 negative and 13 positive (Fig. 3, squares containing − viruses (25, 26), and, in some cases, due to age-dependent contact or +, respectively). The estimated cross-correlations fall outside patterns driven by extensive mixing of children in daycare centers the 2.5% and 97.5% quantile intervals of correlation distribu- and schools (27, 28). The remaining unexplained variation in- tions generated under the null hypothesis of no interaction (see cludes temporal autocorrelations and dependencies between SI Appendix, Tables S1 and S2 and Methods for details). How- viruses. Modeling temporal autocorrelation through a hierarchical

2of9 | www.pnas.org/cgi/doi/10.1073/pnas.1911083116 Nickbakhsh et al. Downloaded by guest on September 26, 2021 Within-Host Virus Mixing Patterns Are Nonrandomly Distributed across A ρ= -0.38 IAV – RV the Patient Population, Indicating Virus Virus Interactions Operate at the Scale of Individual Hosts. To infer virus–virus interactions at the 30 level of individual hosts, we applied multivariable binary logistic 20 regression to the diagnostic records of virus-positive patients. We designed our analysis to eliminate the influence of Berkson’sbias, 10 which can lead to spuriously large or small odds ratios (ORs) when

Prevalence (%) 0 inferring disease–disease associations from hospital-based case- control data (33). To account for any influence of this potential B ρ= -0.48 IBV selection bias, we restricted our analysis to the virus-positive pa- AdV tient subset (see Methods for further details). 15 We infer signatures of virus–virus interactions from the nonrandom patterns of virus mixing captured by coinfection in- 10 formation by assessing whether the propensity of a given virus X 5 to coinfect with another virus Y was higher, lower, or equal to the overall propensity of any (remaining) virus group to coinfect with

Prevalence (%) 0 virus Y. We adjusted for the effects of age, gender, patient origin (hospital versus general practice), and the time period (with re- C ρ= 0.57 spect to the 3 major waves of the 2009 IAV pandemic). To dis- RSV tinguish interactions between explanatory and response viruses 20 MPV from unrelated seasonal changes in infection risk, we also ad- 15 justed for the monthly background prevalence of response virus 10 infections. As our data did not allow us to infer the directionality of virus– 5 virus interactions, and nor did we have an a priori basis to inform Prevalence (%) 0 this, we first performed 72 statistical tests to evaluate all 36 virus- pair hypotheses in 9 virus models (IAV, IBV, RV, RSV, human

D ρ= 0.50 PIV1 coronaviruses [CoV], AdV, MPV, PIVA [PIV1 and PIV3], and POPULATION BIOLOGY 4 PIV2 PIVB [PIV2 and PIV4]; see Table 1 for details). Due to com- 3 paratively low infection frequencies, PIVs were regrouped into PIVA (human respiroviruses) and PIVB (human rubulaviruses). 2 1 0 Prevalence (%) Negative adjusted correlation Positive adjusted correlation

1-Jan-05 1-Jan-06 1-Jan-071-Jan-081-Jan-091-Jan-10 1-Jan-111-Jan-12 1-Jan-13 - Negative unadjusted correlation Year + Positive unadjusted correlation

Fig. 2. Comparative prevalences of viral infections detected among patients in Glasgow, United Kingdom, 2005 to 2013. Prevalence was measured AdV + - - as the proportion of patients testing positive to a given virus among those CoV + + + - - + - tested in each month. (A and B) Asynchronous seasonality, explained by negative epidemiological interactions. (C and D) Synchronous seasonality, IAV + + + - + - explained by positive epidemiological interactions. ρ = significant correlation coefficients from Bayesian multivariate disease mapping analysis of IBV + + + - viral infection time series shown in Fig. 3. See Table 1 for a full description of the viruses. MPV + + + - - + PIV1 - - - - + autoregressive model (32), we were able to directly estimate the PIV2 + - between-virus correlation matrix adjusted for other key alternative PIV3 + + - - - drivers of infection. This bespoke approach revealed many fewer statistically sup- PIV4 - - - ported epidemiological interactions, with negative interactions RSV + + + + - between IAV and RV and between influenza B virus (IBV) and adenovirus (AdV) (Fig. 3, blue squares), as well as positive in- RV - - - - teractions between RSV and MPV and between PIV1 and PIV2 AdV CoV IAV IBV MPV PIV1 PIV2 PIV3 PIV4 RSV RV (Fig. 3, red squares) (SI Appendix, Tables S3 and S4). These A interactions can be seen empirically as asynchronous (Fig. 2 Fig. 3. Negative and positive interactions among influenza and non- and B) and synchronous (Fig. 2 C and D) prevalence trends in influenza viruses at population scale. Significant unadjusted correlations from the raw data. We did not detect epidemiological interactions bivariate cross-correlation analysis applying Spearman’srankmethodto among other possible virus pairs. We note the lack of interaction monthly viral infection prevalences are shown in gray, with negative and − + between the genetically related but antigenically dissimilar pair positive correlations indicated by and , respectively, and noncorrelated virus pairs in white. Significant support for virus–virus interactions based on corre- IAV/IBV, in line with inconsistent patterns of infection asyn- lations derived from Bayesian disease mapping analysis adjusting for fluctua- chrony over the 9-y time frame of our study (SI Appendix, Fig. tions in testing frequency, temporal autocorrelation, and alternative drivers of S1). See Methods for further details. correlated seasonality are shown in blue (negative) and red (positive).

Nickbakhsh et al. PNAS Latest Articles | 3of9 Downloaded by guest on September 26, 2021 Table 1. Details of virus detections by multiplex real-time RT- Table S18 and Methods for details). Our simulations show that a PCR assays temporary reduction in host susceptibility to a secondary com- “ ” Virus nomenclature according to the mon cold-like virus infection, a refractory period, caused by a Abbreviation International Committee on Taxonomy of Viruses primary influenza-like virus infection is sufficient to produce an observable epidemiological impact. For example, a refractory RV Rhinoviruses (A–C) period of just 2 d caused a 23% decrease in common cold-like IAV Influenza A virus* virus incidence during peak influenza-like virus activity, while a IBV Influenza B virus refractory period of 7 d generated as much as a 61% decrease in † RSV (Formerly) respiratory syncytial virus common cold-like virus incidence (Fig. 5A). Notably, these CoV Human coronaviruses (229E, NL63, HKU1) simulations produced asynchronous temporal patterns of in- AdV (Formerly) human adenoviruses‡ fection qualitatively similar to our empirical observations, such MPV Human metapneumovirus that the periodic decline in common cold-like virus infections PIV3 (Formerly) human parainfluenza 3 virus§ coincides with peak influenza-like virus activity (Fig. 5B). PIV1 (Formerly) human parainfluenza 1 virus§ PIV4 (Formerly) human parainfluenza 4 virus¶ Discussion PIV2 (Formerly) human parainfluenza 2 virus¶ In this study, we demonstrated the presence, and examined the nature, of virus–virus interactions at both the epidemiological *A generic assay detecting seasonal H3N2 and H1N1 subtypes and one specific to A(H1N1)pdm09. and individual host level by examining the diagnostic results of † (Currently) human orthopneumovirus. patients that were simultaneously tested for 11 respiratory viruses. ‡ (Currently) human mastadenoviruses (A–G). Our study provides the most comprehensive quantitative support §(Currently) human respiroviruses (1 and 3). to date for the existence of interactions among taxonomically ¶(Currently) human rubulaviruses (2 and 4). broad groups of respiratory viruses, building on earlier work on pathogen–pathogen interactions in the context of wildlife (37), childhood diseases (1, 38), and community ecology (39, 40). We Of the 72 pairwise tests, 17 yielded ORs with P < 0.05 before reveal statistical support for the existence of both positive and correction for multiple testing: 4 negative interactions (OR < 1) negative interspecific interactions among respiratory viruses at among 3 virus pairs involving influenza and 13 positive interactions both population and individual host scales. Our findings extend the (OR > 1) among 8 pairs of noninfluenza viruses (Fig. 4A) (note that established paradigms that have focused on bacteria–bacteria and not all significant interactions were bidirectional). We then applied bacteria–virus relationships (5) to the virus–virus case and dem- Holm’s method (34) to control the familywise error rate. This onstrate that respiratory virus interactions are not only restricted to confirmed the existence of 2 within-host interactions: a negative previously described competitive interference forms (21). interactionbetweenIAVandRV(OR= 0.27, P < 0.001) and a By studying the coinfection patterns of individual patients, our positiveinteractionbetweenAdVandPIVB(OR= 3.99, P < analyses support an interference between influenza and non- 0.001). See SI Appendix,TablesS15–S17 and Methods for details. influenza viruses operating at the host scale. This finding supports We also used a permutation method to test the global null the potential role of innate immunity, such as through the antiviral hypothesis that there were no interactions among any of the action of IFN, as early discussions on virus–virus interactions remaining 5 virus groups (IBV, CoV, MPV, RSV, and PIVA). The had postulated (25, 41). Capturing this potentially immune- distribution of P values tended significantly toward zero, providing mediated interference in mathematical simulations representing strong evidence for the existence of further interactions (P = the cocirculation of a seasonal influenza-like virus and a ubiquitous 0.0021; Fig. 4B). This indication of the likely existence of addi- common cold-like virus, we demonstrated that a short-lived pro- tional virus–virus interactions, that lie beyond the power permitted tective effect, such as that induced by IFN (25), is sufficient to in- by our data, is consistent with the large number of candidate induce the observed asynchronous seasonal patterns we observe for teractions (17 virus pairs in total) that yielded P < 0.05. Only IAV and RV (Fig. 2A). Many factors could contribute to interfer- 3 interactions (95% CI = 0 to 8) were expected when applying the ences observed at the population scale through the removal of permutation test of the global null hypothesis to all 72 tests. See SI susceptible hosts (1, 38). Such effects will likely act on a timescale Appendix, Figs. S2 and S3 and Methods for further details. (on the order of days to weeks) that is similar to our proposed biological mechanism and might therefore act alternatively or in Transient Immune-Mediated Cross-Protection Can Generate Linked tandem to generate epidemiological interactions. Asynchronous Transmission Dynamics of Influenza and RV. Our sta- We detected a further negative interaction at the population tistical analyses provide strong support for a negative interaction level between IBV/AdV which was not corroborated at the indi- between seasonal IAV and the relatively ubiquitous RV, at both vidual host scale. While IBV has a (albeit inconsistent) seasonal population and individual host scales. This negative interaction pattern, typically peaking in winter months, AdV typically peaks may be driven by virus competition for susceptible cells, for ex- around May. However, because our Bayesian hierarchical model ample as a consequence of influenza-induced destruction of cell- adjusts for virus seasonality on a month-by-month basis, it is not surface receptors (35) and/or cell death (36), or as a consequence seasonal differences that explain the negative relationship between of virus-induced innate immune responses, such as the secretion this virus pair. In the absence of a seasonal driver or a host-scale of interferon (IFN), which can cause noninfected neighboring mechanism, it is possible that the lack of cooccurrence of IBV and cells to adopt a protective antiviral state (23, 24). Such biological AdV is explained by other ecological drivers. For example, con- mechanisms would render the host resistant, or only partially valescence or hospitalization induced by one virus may reduce the susceptible, to subsequent viral infection. This prompted us to susceptible pool at risk of exposure to other viruses, as previously ask whether a short-lived, host-scale phenomenon could explain discussed by others in the context of childhood diseases (1, 38). the prominent declines in the prevalence of RV among the pa- Both IAV and IBV viruses exhibited only negative interactions tient population during peak influenza activity (Fig. 2A). at both host and population levels, although the specifics dif- To address this question, we performed epidemiological sim- fered. That they differ in their exact pairwise interactions is ulations of the cocirculatory transmission dynamics of a seasonal unsurprising when considering that these viruses are antigenically influenza-like virus, such as IAV, and a nonseasonal common distinct, constitute different taxonomical genera, and exhibit dif- cold-like virus, such as RV, using ordinary differential equation ferent viral evolutionary rates (20, 42), as well as differences in (ODE) mathematical modeling (see SI Appendix, Fig. S4 and their respective age distributions of infection and some aspects of

4of9 | www.pnas.org/cgi/doi/10.1073/pnas.1911083116 Nickbakhsh et al. Downloaded by guest on September 26, 2021 clinical presentation (43, 44). Furthermore, these viruses exhibit inconsistent patterns of cocirculation (45–47) (see also SI Ap- pendix,Fig.S1) and thus their cooccurrence with other respiratory viruses is expected to vary. Based on these differences between Negative interactions Positive interactions IAV and IBV, it is feasible that their ecological relationships with A (corrected p<0.001) (corrected p<0.001) other viruses have evolved differently. Of further note is the lack of interaction detected between Negative interactions Positive interactions IAV and IBV, since there is some suggestion from global data of (uncorrected p<0.05) (uncorrected p<0.05) a short lag between their outbreak peaks. However, epidemiological data are inconsistent in that they report both asynchrony RV and codominance (46, 47). We believe that a lack of confirmation CoV of interference between IAV and IBV is consistent with current virological understanding. These viruses are genetically and anti- IBV genically different; although studies have identified cross-reactive B cell and T cell epitopes in conserved regions of the viral genome (48), sufficient natural cross-protection has not been demon- MPV strated (48, 49). It is, however, possible that their ecological relationship depends on the particular strains cocirculating. Further work is needed to better understand how the existence and nature IAV of virus–virus interactions varies at the level of virus strains. On the other hand, some evidence exists in support of immune-driven RSV interference between H1N1 and H3N2 subtypes of influenza A (46, 47). Our data did not permit reliable analysis at this level of virus differentiation because low and inconsistent numbers of in- PIVB fluenza cases were routinely subtyped. Age patterns of infection may provide important insights into the AdV opportunity for, or alternatively the consequences of, virus–virus PIVA interactions. A lag in epidemic peaks across children and adults has been observed in the case of RSV (50, 51). Such a lag between ages POPULATION BIOLOGY may influence the potential for interaction with other cocirculating viruses, or it may reflect niche segregation as a consequence of viral B Observed interference. Although an interference between RSV and IAV has Expected been proposed (9, 11, 48), a hypothesis recently supported in an 1e−06 Permuted experimental ferret model (21), this was not supported by our data. It is possible that such an effect was masked because our analysis did not enable stratification of effects by age, a step that is anticipated to 1e−05 severely limit the statistical power of our highly comprehensive analysis of virus–virus interactions. More detailed investigation to 1e−04 address more precisely when, and among whom, virus–virus in-

alue teractions occur will be an important challenge for future research. Our study describes positive interactions among respiratory 0.001 viruses at the population scale. These positive epidemiological interactions were not mirrored at the host scale, which suggests they are independent of host-scale factors and may instead be 0.01 explained by variables that were not captured by our study. For ex-

Observed P−v Observed ample, some respiratory viruses, such as RSV and MPV, are known 0.05 0.1 to enhance the incidence of pneumococcal pneumonia (6, 52). It is possible that we are observing an intermediary effect of secondary 0.5 bacterial infections, which may lead to enhanced hospitalization with 1 both viruses, rather than a direct virus–virus interaction. Although not mirrored at the population level, our 72 pairwise 100.5 .10.05 tests did detect the existence of positive interactions among 8 pairs Expected P−value of viruses, including strong statistical support for AdV/PIVB following Holm’s correction for multiple comparisons. This finding is Fig. 4. Host-scale interactions among influenza and noninfluenza viruses. (A) Statistically supported negative (OR < 1) and positive (OR > 1) virus–virus consistent with a recent, smaller-scale clinical study of children interactions based on uncorrected P < 0.05 from multivariable logistic re- diagnosed with pneumonia, which detected 2 pairs of positively gression analysis. Line widths are proportional to the absolute value of the associated noninfluenza viruses (17). Plausible mechanisms of maximum log OR estimated per virus pair. Two interactions (RV/IAV and positive virus–virus interactions include enhanced viral growth AdV/PIVB) retained strong statistical support (P < 0.001) following Holm’s through syncytia formation of a coinfecting virus (22), the virus- correction to control the familywise error rate. (B) Test of the global null mediated down-regulation of IFN, which may promote invasion by hypothesis: QQ plot of the observed P value distribution from 20 pairwise opportunistic viruses (41), and the virus-induced release of cyto- tests among the 5 remaining virus groups (IBV, CoV, MPV, RSV, and PIVA; kines such as interleukin 10, which prevents the detrimental effects green line), compared to the P value distribution expected under the global of dysregulated immune responses and might lead to immuno- null hypothesis of no interactions (purple dashed line). The distribution of QQ lines simulated from the global null hypothesis using 10,000 permuta- suppression and enhanced susceptibility to secondary viral infections is shown in gray. See Table 1 for a full description of the viruses. Due to tions (53, 54). That most interactions detected at the host scale comparatively low infection frequencies, parainfluenza viruses were regrou- were not supported at the population level is not surprising given ped into PIVA (PIV1 and PIV3; human respiroviruses) and PIVB (PIV2 and PIV4; that interaction effects are reliant on coinfection, or sequen- human rubulaviruses). tial infections, occurring within a short time frame. The relative

Nickbakhsh et al. PNAS Latest Articles | 5of9 Downloaded by guest on September 26, 2021 1 0 cold virus incidence during cations for the improved design of disease forecasting models % decrease in common

A peak inﬂuenza activity and the evaluation of disease control interventions. Future ex- 0.8 −50 perimental studies are required to decipher the biological mechanisms that underpin virus–virus interactions and their effects on the within-host dynamics of infection. 0.6 −100 Methods 0.4 −150 Study Population and Dataset. Our study was based on routine diagnostic test data used to inform the laboratory-based surveillance of acute respiratory

Interaction (φ) infections in NHS Greater Glasgow and Clyde (the largest Health Board in 0.2 −200 Scotland), spanning primary, secondary, and tertiary healthcare settings. Clinical specimens were submitted to the West of Scotland Specialist Virology 0 Centre for virological testing by multiplex real-time RT-PCR (58, 59). Patients 5 10 15 20 25 were tested for 11 groups of respiratory viruses summarized in Table 1. Refractory period (days) Data were available to our study for the period January 2005 to December 2013 comprising 61,427 clinical samples, most of which were nasal and/or throat swabs (98%). The test results of individual samples were aggregated to the patient level using a window of 30 d to define a single episode of illness, B Inﬂuenza-like virus giving an overall infection status per episode of respiratory illness. This yielded a Common cold-like virus: with interference total of 44,230 episodes of respiratory illness from 36,157 individual patients. In Common cold-like virus: without interference this study population, 35% (15,302) of patient episodes were virus-positive, 8% 4 (1247) of which were coinfected with more than one virus group. These data provide a coherent source of routine laboratory-based data for 3 inferring epidemiological patterns of respiratory illness, reflecting typical community-acquired respiratory virus infections in a large urban population (60). Of the total 44,230 episodes of illness, 62% (27,284) were tested for all 2 11 virus groups, increasing to 83% (22,420) among 26,974 episodes tested out with the 3 major waves of A(H1N1)pdm09 virus circulation. More than

% infected 1 99% of patients were tested for all viruses in each calendar month (except for AdV in 2005 to 2007, PIV4 in 2005, and during pandemic influenza). Virological diagnostic assays remained consistent over the 9-y period, with 0 10 11 12 13 14 15 the exception of the RV assay, which was modified during 2009 to detect a Time (years) wider array of RV and enteroviruses (including D68), and 1 of 4 CoV assays (CoV-HKU1) was discontinued in 2012. Fig. 5. Mathematical ODE models simulating the impact of viral in- These diagnostic data included test-negative results providing the nec- terference on the cocirculatory dynamics of a seasonal influenza-like virus essary denominator data to account for fluctuations in testing frequencies and a ubiquitous common cold-like virus. (A) Percentage decrease in the across patient groups and over time. Such changes in testing patterns may be minimum daily incidence of common cold-like virus infections during peak expected to arise in healthcare data accrued over extended time frames, for influenza-like virus activity for varying interaction strengths and refractory example because of changes in healthcare-seeking behavior and/or the periods. (B) Asynchronous incidences of influenza-like virus (in red) and clinical management of patients. We refer readers to ref. 60 for further common cold-like infections in the presence (blue) and absence (green) of details of the dataset and epidemiological patterns of virus detections. interference with the influenza like virus. This example assumes a strong interaction (φ = 1) and 7-d refractory period shown over 10 simulated years. Bivariate Cross-Correlation Analysis. Spearman’s rank correlation coefficients

The R0s of these viruses assuming a completely susceptible homogeneous were computed and tested between all pairs of virus infection prevalences population are 1.6 (virus 1) and 2 (virus 2). The model supports the hy- (the proportion positive among those tested) in each month using cor.test in pothesis that temporary nonspecific protection elicited by influenza explains R (32) version 3.4.4 (61). These analyses were based on 26,974 patient epi- the periodic decline in rhinovirus frequency during peak influenza activity sodes of respiratory illness excluding the period spanning the 3 major waves (Fig. 2A). of A(H1N1)pdm09 virus circulation. We additionally conducted an analysis of the distribution of correlation coefficients generated under the null hypothesis of no virus–virus interactions. To do so, we randomly permuted the rareness of interaction events might thus limit their de- monthly prevalence time series of each virus pair 1,000 times and computed tectability and epidemiological impact. the 2.5% and 97.5% quantiles of each distribution of correlation coefficients. See SI Appendix, Tables S1 and S2 for the estimated correlation co- It is important to consider that virus–virus interactions are likely efficients, distributions under the null hypothesis, and P values. shaped by evolutionary processes and thus their existence, magni- tude, and dominant form (whether competitive or facilitative) may Multivariate Bayesian Hierarchical Models. Although the simple bivariate not be static in nature, and instead depend on particular virus cross-correlation analysis revealed several significantly correlated virus pairs, strains, as well as host and environmental factors (55). The virus– inferring genuine virus–virus interactions using diagnostic patient data re- virus interactions we detect may represent a “net effect” of the quires several features to be accounted for: Absolute infection counts must evolutionary processes (55) driving the virus strains circulating in be adjusted for fluctuations in testing frequency, temporal autocorrelation, the UK population over the 9-y timeframe of this study. It should and alternative confounding explanations of temporal dependency between virus groups. also be borne in mind that a large proportion of respiratory infec- To address these methodological limitations, we developed and applied a tions, including influenza, are expected to be asymptomatic (56), statistical approach that extends a multivariate Bayesian hierarchical mod- and coinfections of some viruses may be associated with attenuated eling method to times-series data (32). The method employs Poisson re- disease (23, 57). It is therefore conceivable that the form of in- gression to model observed monthly infection counts adjusting for teraction detected in a patient population, although of clinical im- confounding covariates and underlying test frequencies. This framework portance, may differ from that occurring in the community at large. uses a multivariate normal distribution to model the random temporal Our study provides strong statistical support for the existence variation in the adjusted virus counts, naturally incorporating a between- virus covariance matrix that enabled estimation of virus–virus correlations. of interactions among genetically broad groups of respiratory Through estimating, and scaling, the off-diagonal entries of this matrix, we viruses at both population and individual host scales. Our find- were able to estimate posterior interval estimates for correlations between ings imply that the incidence of influenza infections is interlinked each virus pair. Under a Bayesian framework, posterior probabilities were with the incidence of noninfluenza viral infections with impli- estimated to assess the probability of zero being included in each interval

6of9 | www.pnas.org/cgi/doi/10.1073/pnas.1911083116 Nickbakhsh et al. Downloaded by guest on September 26, 2021 (one for each virus pair). Adjusting for multiple comparisons, correlations area under the receiver operator characteristic curve. Because these data do corresponding to intervals with an adjusted probability less than 0.05 were not allow for inferences of the directionality of the virus–virus interactions, deemed significantly different from zero. Crucially, the method makes use of we conducted all possible 72 pairwise statistical tests in the first instance to multiple years of data, allowing expected annual patterns for any virus to be evaluate 36 virus-pair hypotheses and present results where ORs yield un- estimated, thereby accounting for typical seasonal variability in infection risk adjusted P < 0.05. These analyses detected 17 significant virus–virus in- while also accounting for covariates such as patient age (as well as gender and teraction (4 negative and 13 positive) out of the 72 tests (SI Appendix, Tables hospital vs. general practice [GP] patient origin). We identified significant in- S15 and S16). Not all virus-pair interactions were bidirectional; however, we teractions between IAV/RV, IBV/AdV, RSV/MPV, and PIV1/PIV2. See SI Appen- note that these analyses do not allow the differentiation of cause dix,TablesS3andS4for the pairwise correlation estimates summarized in Fig. 3 from effect. of the main text, and see ref. 32 for further details on the method itself. Applying Holm’s method to control the probability of one or more false discoveries arising among the family of 72 tests [using the p.adjust function Host-Scale Analyses: Binary Logistic Regression. We analyzed the patient-level in R version 3.4.4 (61)] yielded strong evidence (P < 0.001 following correc- coinfection data with a series of binary logistic regression models and tion) of the existence of both negative (IAV/RV) and positive (AdV/PIVB) evaluate whether virus–virus interactions may be explained by interactions interactions (see SI Appendix, Table S17 for details). A permutation test of operating at the scale of individual hosts. Of the 44,230 total episodes of the global null hypothesis was then applied to the 5 remaining virus groups respiratory illness, 73% represent patients experiencing a single episode dur- (IBV, CoV, MPV, RSV, and PIVA) to test the hypothesis that the 20 remaining ing the 9-y study period, while the remaining 27% represent patients expe- null hypotheses tested were true. Using Fisher’s method of combining P riencing2ormoreepisodes(range2to26;mean= 3 episodes per patient). values (62), we found strong evidence of association among these 5 virus Retaining the first observed episodes generated 36,157 patient records and groups (P = 0.0021; SI Appendix,Fig.S2), although we expect nonindependence 20,703 with complete testing across all virus groups [note that pressures on between these tests. We therefore accounted for nonindependence laboratory resources led to high partial testing during the major waves of among the pairwise tests by using permutations to simulate the null dis- A(H1N1)pdm09 virus circulation]. Of these 20,703 patients, 6,884 were virus- tribution of combined P values. Each generalized linear model was fitted to positive, 11% of which were coinfected with 2 or more viruses. We note that 10,000 datasets where the null hypothesis was simulated by permuting the the coinfection prevalence and detected virus–virus interactions were robust response variable (virus Y). The significant tendency of the distribution of P to the choice of window used to define an episode of illness when aggre- values toward zero (and away from the uniform distribution expected gating individual samples (see SI Appendix, Table S19 for details). under the null hypothesis) was illustrated using a quantile–quantile (QQ) We designed our analysis to minimize the potential influence of Berkson’s plot to compare the 20 observed P values with the 10,000 distributions bias—a form of selection bias arising in studies aiming to infer disease–disease simulated under the global null hypothesis (Fig. 3C). The signal of addi- associations from hospital-based case-control data. This bias arises where there tional interactions was further demonstrated when the permutation test is an underlying difference in the probabilities of study inclusion between case of the global null hypothesis was extended to all 72 tests (SI Appendix, and control groups (33). Our approach was, for each study virus (Y) in turn, to Fig. S3)—only 3 (95% CI: 0 to 9) interactions were expected by chance

restrict the analysis to patients positive for one of the other (explanatory) assuming the null hypothesis of no interactions, in comparison to the POPULATION BIOLOGY viruses, thereby eliminating the influence of the underrepresentation of the 17 interactions we detected. virus-negative “healthy” population, implicit in the patient-based study population. The study population comprised individuals infected with at least one Mathematical Modeling of Influenza and RV Interactions and Population other (non-Y) virus. Within that group, exposed individuals were positive Impact. We developed a 2-pathogen deterministic SIR-type mechanistic to virus X, and unexposed individuals were negative to virus X. Cases were model to study the population dynamics of a seasonal influenza-like virus coinfected with virus Y, while controls were negative to virus Y. In this and a ubiquitous common cold-like virus cocirculation. We used this way, our analysis quantifies whether the propensity of virus X to coinfect framework to compare the frequency of common cold-like virus infections with virus Y was more, less, or equal to the overall propensity of any (remaining) with and without an interference with the influenza-like virus. A schematic virus group to coinfect with Y. Our study therefore infers signatures of virus– representation of the model is provided in SI Appendix,Fig.S4.The virus interactions from the nonrandom patterns of virus mixing among the virus- temporal dynamics of the viruses were distinguished in 2 key ways. First, positive population. seasonal forcing was applied to the influenza-like virus (virus 1) via a si- Our analyses adjusted for key predictors of respiratory virus infections: nusoidally varying transmission rate. Second, the rate of waning immunity patient age (AGE.CAT), patient sex (SEX), hospital vs. GP patient origin of the common cold-like virus (virus 2) was assumed to be 10 times faster (ORIGIN), and time period of sample collection with respect to the influenza than for the influenza-like virus. This more rapid replenishment of sus- A(H1N1)pdm09 virus pandemic (PANDEMIC). In addition, we ensured that ceptible individuals was designed to reflect the high year-round preva- virus–virus associations were not simply explained by independent monthly lence and diversity of circulating subtypes that are characteristic of RV fluctuations in the response virus infection risk. To do so, we adjusted the infections (63). total number of infections with the response virus (VCOUNT) and the total Individuals were assumed to acquire primary infections at rate β (days−1), number tested (TCOUNT) within a 15-d window either side of each (earliest) assuming frequency-dependent transmission, and subsequently transition sample collection date for each individual observation. Model variables were from susceptible (S) to an infectious refractory phase (I). Infected individ- either binary or categorized; see SI Appendix, Table S5 for details. uals were assumed not to be susceptible to further infections with the Specifically, the relative odds of coinfection with virus Y (versus any primary infecting virus. Our assumption is that multiple exposures to other virus group) was estimated for each of the 8 explanatory viruses, for similar virus strains are unlikely to alter the within-host dynamics during each response virus Y. Thus, each virus model (IAV, IBV, RV, RSV, CoV, AdV, this short period. At the end of the infectious period, individuals entered a − MPV, PIVA, and PIVB) used a specific subset of the virus-positive population noninfectious refractory phase (J) at rate 1/γ (days 1). This second re- (because each virus model excluded virus-negative patients and single fractory phase was designed to reflect immune effects that may persist for infections involving only the response virus Y) with each data subset a period beyond viral clearance (64, 65). During both refractory phases, ranging from 4,629 to 6,743 (see SI Appendix, Tables S6–S14 for details). viral interactions are captured via reduced susceptibility of influenza-like Due to low infection frequencies, we regrouped PIVs into 2 groups: PIVA virus infected individuals to either coinfection with the common cold-like (PIV1 and PIV3; human respiroviruses) and PIVB (PIV2 and PIV4; human virus (during phase I) or, alternatively, a secondary infection with the rubulaviruses). common cold-like virus (during phase J). The strength of the interaction φ Each of the 9 virus models followed the following form, where PY indicates was determined by an interaction parameter applied to the transmission i φ = the probability that individual i is coinfected with virus Y, and X1...X8in- rate, whereby 0 induced complete susceptibility (equivalent to trans- < φ < dicates the 8 explanatory viruses: mission dynamics in the absence of an interaction), 0 1 induced a φ = ! partial reduction in susceptibility, and 1 induced a complete loss PY of susceptibility. log i = β + β AGE. CAT + β SEX + β ORIGIN + β PANDEMIC 1 − PY 0 1 i 2 i 3 i 4 i At the end of the second refractory phase, infected individuals developed i long-lasting immunity to infection due to antibodies generated at rate φ + β + + β + + β + β + β −1 5 logðV. COUNTi 1Þ 6 logðT. COUNTi 1Þ 7X1i 8X2i 9X3i (days ) (R). During this phase, individuals were not susceptible to the pri- + β + β + β + β + β mary infection but could acquire secondary infections if previously un- 10X4i 11X5i 12X6i 13X7i 14X8i . exposed. Following a period of long-term immunity individuals reverted to − Analyses were conducted using the glm function in R version 3.4.4 (61). The susceptible (S) at a rate σ (days 1) representing the waning of antibodies quality of each model was assessed by the predictive power given by the against the primary infection.

Nickbakhsh et al. PNAS Latest Articles | 7of9 Downloaded by guest on September 26, 2021 The full model ODEs are given as follows: (the seasonal influenza-like virus) and virus 2 (the nonseasonal common cold- like virus) for each of 12 combinations of infection classes, whereby S = dfSSg ’ ’ = μ + σ fRSg + σ fSRg − β ðtÞfSSgI − β fSSgI − μfSSg = = dt 1 2 1 1 2 2 susceptible, I infectious refractory phase, J noninfectious refractory phase, R = immunity phase, and subscripts 1 and 2 denote the corresponding dfISg ’ ’ virus-specific parameters. For example, {SI} indicates the state of being sus- = β ðtÞfSSgI + σ fIRg − γ fISg − β ð1 − φÞfISgI − μfISg dt 1 1 2 1 2 2 ’ ’ ceptible to virus 1 and infectious with virus 2. I1 and I2 represent the total proportion of individuals infectious with virus 1 and virus 2 respectively; β1(t) dfJSg ’ = γ fISg + σ2fJRg − ω1fJSg − β ð1 − φÞfJSgI − μfJSg denotes the time-dependent rate at which new infections with virus 1 are dt 1 2 2 generated and comprises a fixed transmission parameter B together with a

dfRSg ’ seasonal amplitude parameter b; and φ represents a background, non- = ω fJSg + σ fRRg − σ fRSg − β fRSgI − μfRSg dt 1 2 1 2 2 acute-respiratory-infection–induced rate of mortality given by 1/the average life expectancy (days−1). The peak proportion of individuals dfSIg ’ ’ = σ fRIg + β fSSgI − β ðtÞfSIgI − γ fSIg − μfSIg coinfected with both viruses was 0.39% during each simulated season of dt 1 2 2 1 1 2 influenza-like virus circulation. The R0s of these 2 viruses assuming a

dfIIg ’ ’ completely susceptible homogeneous population are 1.6 (virus 1) and 2 = β ðtÞfSIgI + β ð1 − φÞfISgI − γ fIIg − γ fIIg − μfIIg dt 1 1 2 2 1 2 (virus 2). Full parameter values and ranges are provided in SI Appendix, Table S18. dfJIg = γ + β − φ ’ − ω − γ − μ This framework was implemented in MATLAB software v.R2013b using 1fIIg 2ð1 ÞfJSgI2 1fJIg 2fJIg fJIg dt the ode45 differential equation solver. Each simulation was run for a period of 20 y; a “burn-in” period of 10 y was excluded from the analyses. dfRIg ’ = ω1fJIg + β fRSgI − σ1fRIg − γ fRIg − μfRIg Using this framework, we quantified the effect of transient immune- dt 2 2 2 mediated viral interactions on the percentage decrease in daily nonseasonal

dfSRg ’ common cold-like virus prevalence during peak seasonal influenza-like = σ fRRg + γ fSIg − β ðtÞfSRgI − σ fSRg − μfSRg dt 1 2 1 1 2 virus activity.

dfIRg = β ’ + γ − γ − σ − μ Data Availability. The patient-level data used in this study are available upon 1ðtÞfSRgI1 2fIIg 1fIRg 2fIRg fIRg dt request to NHS Scotland (https://www.informationgovernance.scot.nhs.uk/ dfJRg pbpphsc/home/for-applicants/). Aggregated forms of summary data and = γ fIRg + γ fJIg − ω fJRg − σ fJRg − μfJRg dt 1 2 1 2 computer code may be made available upon request to the corresponding author. dfRRg = ω1fJRg + γ fRIg − σ1fRRg − σ2fRRg − μfRRg dt 2 ACKNOWLEDGMENTS. We thank Bryan Grenfell and Dan Haydon for their critique of the manuscript. This work was supported by the Medical β = + π t Research Council of the United Kingdom (Grant MC_UU_12014/9). L.M. 1ðtÞ B* 1 b*sin 2* * 365 and R.R. are grateful for support from National Science Foundation Grant DEB1216040; Biotechnology and Biological Sciences Research Council ’ = + + I1 IS II IR Grants BB/K01126X/1, BB/L004070/1, BB/L018926/1, BB/N013336/1, BB/ L004828/1, and BB/R012679/1; Foods Standards Agency Grant FS101055; ’ = + + + and the Scottish Government Rural and Environment Science and Analyt- I2 SI II JI RI, ical Services Division, as part of the Centre of Expertise on Animal Disease where {1,2} denotes the infection status of individuals with respect to virus 1 Outbreaks (EPIC).

1. P. Rohani, D. J. Earn, B. Finkenstädt, B. T. Grenfell, Population dynamic interference 16. R. M. Greer et al., Do rhinoviruses reduce the probability of viral co-detection during among childhood diseases. Proc. Biol. Sci. 265, 2033–2041 (1998). acute respiratory tract infections? J. Clin. Virol. 45,10–15 (2009). 2. J. O. Lloyd-Smith, Vacated niches, competitive release and the community ecology of 17. V. G. Nolan et al., Etiology and impact of coinfections in children hospitalized with pathogen eradication. Philos. Trans. R Soc. Lond. B Biol. Sci. 368, 20120150 (2013). community-acquired pneumonia. J. Infect. Dis. 218, 179–188 (2018). 3. S. Shrestha et al., Identifying the interaction between influenza and pneumococcal 18. H. Susi, B. Barrès, P. F. Vale, A. L. Laine, Co-infection alters population dynamics of pneumonia using incidence data. Sci. Transl. Med. 5, 191ra84 (2013). infectious disease. Nat. Commun. 6, 5975 (2015). 4. S. Shrestha, A. A. King, P. Rohani, Statistical inference for multi-pathogen systems. 19. N. Ferguson, R. Anderson, S. Gupta, The effect of antibody-dependent enhancement PLoS Comput. Biol. 7, e1002135 (2011). on the transmission dynamics and persistence of multiple-strain pathogens. Proc. 5. A. A. Bosch, G. Biesbroek, K. Trzcinski, E. A. Sanders, D. Bogaert, Viral and bacterial Natl. Acad. Sci. U.S.A. 96, 790–794 (1999). interactions in the upper respiratory tract. PLoS Pathog. 9, e1003057 (2013). 20. N. M. Ferguson, A. P. Galvani, R. M. Bush, Ecological and immunological determinants 6. D. M. Weinberger, K. P. Klugman, C. A. Steiner, L. Simonsen, C. Viboud, Association of influenza evolution. Nature 422, 428–433 (2003). between respiratory syncytial virus activity and pneumococcal disease in infants: A 21. S. Bhattacharyya, P. H. Gesteland, K. Korgenski, O. N. Bjørnstad, F. R. Adler, Cross- time series analysis of US hospitalization data. PLoS Med. 12, e1001776 (2015). immunity between strains explains the dynamical pattern of paramyxoviruses. Proc. 7. L. Opatowski, M. Baguelin, R. M. Eggo, Influenza interaction with cocirculating Natl. Acad. Sci. U.S.A. 112, 13396–13400 (2015). pathogens and its impact on surveillance, pathogenesis, and epidemic profile: A key 22. M. Shinjoh, K. Omoe, N. Saito, N. Matsuo, K. Nerome, In vitro growth profiles of role for mathematical modelling. PLoS Pathog. 14, e1006770 (2018). respiratory syncytial virus in the presence of influenza virus. Acta Virol. 44,91–97 8. J. A. McCullers, J. E. Rehg, Lethal synergism between influenza virus and Strepto- (2000). coccus pneumoniae: Characterization of a mouse model and the role of platelet- 23. K. F. Chan et al., Investigating viral interference between influenza A virus and hu- activating factor receptor. J. Infect. Dis. 186, 341–350 (2002). man respiratory syncytial virus in a ferret model of infection. J. Infect. Dis. 218, 406– 9. G. Anestad, Interference between outbreaks of respiratory syncytial virus and influ- 417 (2018). enza virus infection. Lancet 1, 502 (1982). 24. A. J. Gonzalez, E. C. Ijezie, O. B. Balemba, T. A. Miura, Attenuation of influenza A 10. F. B. Bang, Epidemiological interference. Int. J. Epidemiol. 4, 337–342 (1975). virus disease severity by viral coinfection in a mouse model. J. Virol. 92, e00881-18 11. W. P. Glezen, A. Paredes, L. H. Taber, Influenza in children. Relationship to other (2018). respiratory agents. JAMA 243, 1345–1349 (1980). 25. A. Isaacs, D. C. Burke, Viral interference and interferon. Br. Med. Bull. 15, 185–188 12. A. Linde, M. Rotzén-Ostlund, B. Zweygberg-Wirgart, S. Rubinova, M. Brytting, Does (1959). viral interference affect spread of influenza? Euro Surveill. 14, 19354 (2009). 26. H. Goto et al., Enhanced growth of influenza A virus by coinfection with human 13. J. S. Casalegno et al., Rhinoviruses delayed the circulation of the pandemic influenza parainfluenza virus type 2. Med. Microbiol. Immunol. (Berl.) 205, 209–218 (2016). A (H1N1) 2009 virus in France. Clin. Microbiol. Infect. 16, 326–329 (2010). 27. C. Viboud et al., Risk factors of influenza transmission in households. Br. J. Gen. Pract. 14. J. S. Casalegno et al., Impact of the 2009 influenza A(H1N1) pandemic wave on the 54, 684–689 (2004). pattern of hibernal respiratory virus epidemics, France, 2009. Euro Surveill. 15, 19485 28. S. Cauchemez, A. J. Valleron, P. Y. Boëlle, A. Flahault, N. M. Ferguson, Estimating the (2010). impact of school closure on influenza transmission from Sentinel data. Nature 452, 15. G. C. Mak, A. H. Wong, W. Y. Ho, W. Lim, The impact of pandemic influenza A (H1N1) 750–754 (2008). 2009 on the circulation of respiratory viruses 2009-2011. Influenza Other Respir. Viruses 29. J. Shaman, M. Kohn, Absolute humidity modulates influenza survival, transmission, 6,e6–e10 (2012). and seasonality. Proc. Natl. Acad. Sci. U.S.A. 106, 3243–3248 (2009).

8of9 | www.pnas.org/cgi/doi/10.1073/pnas.1911083116 Nickbakhsh et al. Downloaded by guest on September 26, 2021 30. Y. G. Karim, M. K. Ijaz, S. A. Sattar, C. M. Johnson-Lussenburg, Effect of relative hu- 49. C. D. O’Donnell et al., Effect of priming with H1N1 influenza viruses of variable an- midity on the airborne survival of rhinovirus-14. Can. J. Microbiol. 31, 1058–1061 tigenic distances on challenge with 2009 pandemic H1N1 virus. J. Virol. 86, 8625–8633 (1985). (2012). 31. V. E. Pitzer et al., Environmental drivers of the spatiotemporal dynamics of respiratory 50. D. M. Fleming et al., Modelling estimates of the burden of Respiratory Syncytial virus syncytial virus in the United States. PLoS Pathog. 11, e1004591 (2015). infection in adults and the elderly in the United Kingdom. BMC Infect. Dis. 15, 443 32. C. Mair et al., Estimation of temporal covariances in pathogen dynamics using (2015). Bayesian multivariate autoregressive models. arXiv:1611.09063v4 (13 April 2019). 51. M. C. Zambon, J. D. Stockton, J. P. Clewley, D. M. Fleming, Contribution of influenza ’ 33. J. Schwartzbaum, A. Ahlbom, M. Feychting, Berkson sbiasreviewed.Eur. J. Epidemiol. and respiratory syncytial virus to community cases of influenza-like illness: An ob- – 18,1109 1112 (2003). servational study. Lancet 358, 1410–1416 (2001). – 34. S. Holm, A simple sequentially rejective multiple test procedure. Scand. J. Stat. 6,65 52. S. A. Madhi, K. P. Klugman; Vaccine Trialist Group, A role for Streptococcus pneumoniae 70 (1979). in virus-associated pneumonia. Nat. Med. 10,811–813 (2004). 35. A. D. Mesecar, K. Ratia, Viral destruction of cell surface receptors. Proc. Natl. Acad. Sci. 53. J. M. Rojas, M. Avia, V. Martin, N. Sevilla, IL-10: A multifunctional cytokine in viral – U.S.A. 105, 8807 8808 (2008). infections. J. Immunol. Res. 2017, 6104054 (2017). 36. V. S. Hinshaw, C. W. Olsen, N. Dybdahl-Sissoko, D. Evans, Apoptosis: A mechanism of 54. S. Davidson, M. K. Maini, A. Wack, Disease-promoting effects of type I interferons in cell killing by influenza A and B viruses. J. Virol. 68, 3667–3673 (1994). viral, bacterial, and coinfections. J. Interferon Cytokine Res. 35, 252–264 (2015). 37. S. Telfer et al., Species interactions in a parasite community drive infection risk in a 55. F. Zélé, S. Magalhães, S. Kéfi, A. B. Duncan, Ecology and evolution of facilitation wildlife population. Science 330, 243–246 (2010). among symbionts. Nat. Commun. 9, 4869 (2018). 38. P. Rohani, C. J. Green, N. B. Mantilla-Beniers, B. T. Grenfell, Ecological interference 56. A. C. Hayward et al.; Flu Watch Group, Comparative community burden and severity between fatal diseases. Nature 422, 885–888 (2003). of seasonal and pandemic influenza: Results of the flu Watch cohort study. Lancet 39. A. B. Pedersen, A. Fenton, Emphasizing the ecology in parasite community ecology. Respir. Med. 2, 445–454 (2014). Trends Ecol. Evol. 22, 133–139 (2007). 57. E. A. Goka, P. J. Vallely, K. J. Mutton, P. E. Klapper, Single and multiple respiratory 40. E. C. Griffiths, A. B. Pedersen, A. Fenton, O. L. Petchey, Analysis of a summary network virus infections and severity of respiratory disease: A systematic review. Paediatr. of co-infection in humans reveals that parasites interact most via shared resources. Respir. Rev. 15, 363–370 (2014). Proc. Biol. Sci. 281, 20132286 (2014). 58. R. N. Gunson, T. C. Collins, W. F. Carman, Real-time RT-PCR detection of 12 respiratory 41. F. Dianzani, Viral interference and interferon. Ric. Clin. Lab. 5, 196–213 (1975). – 42. E. Nobusawa, K. Sato, Comparison of the mutation rates of human influenza A and B viral infections in four triplex reactions. J. Clin. Virol. 33, 341 344 (2005). viruses. J. Virol. 80, 3675–3678 (2006). 59. R. N. Gunson, S. Bennett, A. Maclean, W. F. Carman, Using multiplex real time PCR in – 43. V. Peltola, T. Ziegler, O. Ruuskanen, Influenza A and B virus infections in children. order to streamline a routine diagnostic service. J. Clin. Virol. 43, 372 375 (2008). Clin. Infect. Dis. 36, 299–305 (2003). 60. S. Nickbakhsh et al., Extensive multiplex PCR diagnostics reveal new insights into the – 44. J. Yap et al., Differing clinical characteristics between influenza strains among young epidemiology of viral respiratory infections. Epidemiol. Infect. 144, 2064 2076 (2016). healthy adults in the tropics. BMC Infect. Dis. 12, 12 (2012). 61. R Core Team, R: A Language and environment for statistical computing (R Foundation 45. Y. Iha et al., Comparative epidemiology of influenza A and B viral infection in a for Statistical Computing, Vienna, 2013). ’ – subtropical region: A 7-year surveillance in Okinawa, Japan. BMC Infect. Dis. 16, 650 62. R. C. Elston, On Fisher s method of combining p-values. Biom. J. 33, 339 345 (1991). (2016). 63. A. C. Palmenberg et al., Sequencing and analyses of all known human rhinovirus 46. D. He et al., Global spatio-temporal patterns of influenza in the post-pandemic era. genomes reveal structure and evolution. Science 324,55–59 (2009).

Sci. Rep. 5, 11013 (2015). 64. F. G. Hayden et al., Local and systemic cytokine responses during experimental human POPULATION BIOLOGY 47. B. S. Finkelman et al., Global patterns in seasonal activity of influenza A/H3N2, A/H1N1, influenza A virus infection. Relation to symptom formation and host defense. J. Clin. and B from 1997 to 2005: Viral coexistence and latitudinal gradients. PLoS One 2,e1296 Invest. 101, 643–649 (1998). (2007). 65. D. D. Richman, B. R. Murphy, S. Baron, C. Uhlendorf, Three strains of influenza A virus 48. M. Terajima, J. A. B. Babon, M. D. T. Co, F. A. Ennis, Cross-reactive human B cell and (H3N2): Interferon sensitivity in vitro and interferon production in volunteers. J. Clin. T cell epitopes between influenza A and B viruses. Virol. J. 10, 244 (2013). Microbiol. 3, 223–226 (1976).

Nickbakhsh et al. PNAS Latest Articles | 9of9 Downloaded by guest on September 26, 2021